Fault prognosis using dynamic wavelet neural networks

Abstract

Prognostic algorithms for condition based maintenance of critical machine components are presenting major challenges to software designers and control engineers. Predicting time-to-failure accurately and reliably is absolutely essential if such maintenance practices are to find their way into the industrial floor. Moreover, means are required to assess the performance and effectiveness of these algorithms. This paper introduces a prognostic framework based upon concepts from dynamic wavelet neural networks and virtual sensors and demonstrates its feasibility via a bearing failure example. Statistical methods to assess the performance of prognostic routines are suggested that are intended to assist the user in comparing candidate algorithms. The prognostic and assessment methodology proposed here may be combined with diagnostic and maintenance scheduling methods and implemented on a conventional computing platform to serve the needs of industrial and other critical processes.